Dessa: Block's Full-Duplex Speech Model Trained Only on Audio

Dessa is Block's full-duplex speech model, and the Hacker News reaction split between awe at the audio and alarm at the implications

Dessa is a conversational speech model released by Block (the company behind Square, Cash App, and Afterpay) that is trained only on speech, runs in full duplex so it can speak while you speak, and — according to the announcement as quoted in the discussion — runs locally on a MacBook Pro M4 Max or better. It reached the Hacker News front page with 496 points and 187 comments, where commenters described the demo audio as the most convincing synthetic voice they had heard while also arguing that the technology's most obvious near-term uses are impersonation and fraud.

A note on sourcing: the scraped content of the linked page (block.xyz/dessa) contained only a Block privacy notice and no product description, so the technical claims below come from the Hacker News discussion and from what commenters quoted from the announcement page itself.

What the announcement page reportedly claims

The claim that drew the most attention was about deployment and training cost rather than model architecture. A commenter quoting the announcement summarized it as follows:

The conversation examples are impressive but this made me say holy shit out loud: "it runs locally on a MacBook Pro M4 Max or better. And we figured out how to train Dessa on only one NVIDIA GB200 NVL72." If it really can replicate that low latency + that expressiveness on local devices, that's one heck of an accomplishment for "indie" AI.

Two separate claims are bundled there: efficient local inference on consumer Apple Silicon, and training on a single NVIDIA GB200 NVL72 rack-scale system. Both are unusual for a model whose output is judged lifelike in real time.

The Hacker News submission title was "Dessa: the most authentic full-duplex speech model ever built," while the page itself carried the title "Dessa - the Conversation Model trained only on speech." At least one commenter objected to the superlative framing and to the pattern of submitting sensational titles that are later edited by moderators:

Also, the trend of posting on HN with much more sensationalist title first, getting to the front page, then getting the title adjusted by a moderator should probably be penalized; how is 'most authentic ever' not a marketing term?

Full duplex means the model speaks at the same time you do, not in turns

A full-duplex speech model generates and consumes audio simultaneously, so it can overlap, backchannel, and interrupt rather than waiting for a turn boundary the way a conventional voice assistant pipeline does. Dessa's demos lean hard into that behavior, and that is exactly where the reaction divided.

The strongest endorsement treated the audio as indistinguishable from a person:

ya if this is real this is magnitudes better than anything else out there, it honestly feels like it can't be real because its too real, i would be 100% fooled and never even question it wasn't a real person

Another commenter described the effect the same way:

There's no way I would be able to detect if this was AI from the examples alone. If I got a phone call from a malicious agent using this tech, I probably wouldn't realize it.

The counterargument was that the overlap is overdone — that the model interrupts for the sake of demonstrating that it can:

These bidi/duplex demos always come off as overly artificial to me. It's cool that the model can speak at the same time as you instead of taking turns, but training it to interrupt as much as possible isn't the way to do it. In the very first example, did anyone here find the laugh after the "winter" answer to be convincing? What was the need for it? Why did the human speaker not get to finish a single sentence or finish a thought before the model took over with a monologue?

A related critique held that the conversational content is thin beneath the delivery:

This feels like a parlor trick where the AI was trained to mimic informal telephone mannerisms. The actual conversations were pretty bad. Lots of style, not much substance basically.

Several commenters noted that the human half of the published sample conversations sounded less natural than the model half, and one observed this directly: "I thought it sounded a little stilted and weird and then I realized that was the human half of the conversation."

The training data question: what does a model learn from speech alone?

Because Dessa is trained on audio rather than on text, it is not a speech-to-text step bolted onto an LLM. Commenters argued that this is the source of its expressiveness — and also the source of its limits.

the interesting part to me is that training on voice only is what allows models to learn emotions - voice transmits a lot of emotional information which text doesn't capture. Language has words for a very limited number of emotions - maybe only the peaks - but there's so many more which we can't describe or transmit via text.

A separate thread asked whether a speech-native model actually models sound rather than words:

The real question is, does it actually understand sound? Conventional LLMs are just running speech to text and then working on the text, so they can't understand music or humming or any sort of sound.

Commenters also speculated about the corpus from the demo audio. One guessed podcasts from brief listens; another said the voice "does sound like a human talking but in an affected unnatural way." A specific failure mode was noted: the model referred to the film Memento as "The Mento" before being corrected, which prompted a question about what tokenization a speech-only model uses.

The complaint that the model is fluent but not knowledgeable recurred across the thread. One commenter put it as "lots of style, not much substance"; another described the output as sounding "like a teenager's stream of consciousness." Another framed the tradeoff as intentional: "It doesn't need to be smart the way a SOTA LLM is smart. There's tons of value here already. You can tack it onto another system for deep intelligence."

Compute efficiency: is one GB200 NVL72 the interesting number?

The training-cost claim prompted the thread's most forward-looking question — whether the same efficiency could transfer to text models, which typically consume far more compute:

"we figured out how to train Dessa on only one NVIDIA GB200 NVL72." This is the crazy part. It seems pretty smart, and LLMs require 1000x more compute to train. Can this kind of efficiency somehow be achieved with text too?

No answer was established in the discussion. A related observation pointed out that language models must acquire far more than conversational English: "LLMs need an insane amount of text. Dessa learned language from hundreds of times fewer words (in fairness, LLMs also have to learn C++ and whatever the hell Angular is)."

The integration gap: tool calling, agents, and access

If Dessa is a speech-to-speech model, how does it act on the world? A commenter flagged tool calling as the practical dividing line between demos and agents, arguing that gpt-live-1 is notable precisely because it can be used as a practical agent.

One issue that the full duplex models always have is tool calling. This is where GPT-Live shines in the sense that you can use it as a fully practical agent. Looking forward to seeing how Dessa folks implement this.

Others asked directly how the model reaches anything that is not an audio channel: "If you actually want your 'AI' to interact with anything digital, are these voice-to-voice models going to...use speech to text? How does this model interact with anything that isn't an audio channel?" A related commenter suspected the speech-to-speech design is bound to whatever it was trained with: "Would be nice if you could pipe it into a proper brain, but I am assuming the speech2speech means it's fixed with whatever it was trained with."

Availability was a recurring complaint. The most direct version:

Open weights or it didn't happen.

Commenters asked whether there is any way to test it online, whether it works beyond English (one Greek speaker asked specifically whether their language is supported), and whether it can be run in llama.cpp locally. None of those questions were answered in the thread as captured.

The dominant concern: impersonation, scams, and social isolation

The single most common criticism was not about audio quality but about what authentic-sounding real-time speech enables. The most concrete scenario described in the thread was social engineering against people who know the target's voice:

Even worse, someone could record our voice and fine-tune a model like this so that it can produce our voice and then talk to our family members to deceive them. It'd be a real trouble.

One scenario: An AI voice trainer trains on someone's social media video posts and get a pretty perfect copy. Then SIM jacks the target's phone number and calls pretending to need help. AIs could be used to mock ransom people, request bail money, ask for a change on a production server, etc.

Others focused on the human cost of the conversational realism itself:

I do worry with this kind of fidelity to a human interaction that people will get more drawn into AI relationships and replace human contact... This kind of fidelity blurs the lines of who or what the human is talking to and lets humans fall into an isolated world where they believe they are "socializing."

Very technically impressive, but the fake conversation with the machine was cringe worthy to me; this pretending to care doesn't make me want to converse at all.

The demo set includes an AI-companion example — one commenter cited "Hey, baby" at the start of the third example and reacted with "Block is in the AI girlfriend market? Ew." Another commenter observed that the AI-girlfriend use case appeared to be "the 3rd example" in a list that also includes a conversation between ChatGPT and Dessa.

One commenter argued the whole category is best understood as impersonation technology rather than as a reasoning aid: "Most recent voice-related tech is using voice as an interface or interaction mode... This model is clearly not so smart. So the voice is the whole point, there's not much value beyond it, at least on its own. What that adds up to is, impersonation."

Skepticism about the timeline was also voiced. One commenter estimated that a voice interface could pass a five-minute Turing test within five to ten years, arguing that current LLMs are "an order of magnitude too slow for convincing interactive latency" without extended reasoning time, while conceding that "systems are good enough today for lots of applications."

Where commenters saw legitimate value

Not every use case discussed was adversarial. The most developed constructive argument was language learning:

This is extremely valuable for language education. This is like having on demand access to a conversational partner at all times. But there is no embarrassment factor, patience, or concern about wasting an instructors time. Because the model is trained entirely on audio and not a voice to text add on, you can provide a human extremely high quality feedback and training. And it's cheap to run.

A long comment imagined expressive, interactive audiobooks — narrators that pause when you interrupt, react to questions about a character, adjust volume when you fall asleep, remember your preferences years later — plus game narration and persistent in-game companions.

Others flagged the historical echo of Google Duplex, the 2018 system that booked appointments by phone, and one commenter asked how Google achieved that effect years earlier.

What the discussion did not establish

Several basic facts were never settled in the thread: the model's architecture and parameter count, its measured latency, whether weights or an API would be released, which languages it supports, and how it would call tools or connect to non-audio systems. The complaints about content quality — "lots of style, not much substance" — and the enthusiasm about voice quality — "I would be 100% fooled" — are both judgments of the same published demo tracks, and neither side disputed that the audio fidelity itself is a step change for full-duplex speech.